From copilots to agents: where supervised AI is starting to pay in bank operations

For most of 2023 and 2024, generative AI in banking meant copilots: assistants that drafted emails, summarised documents and answered policy questions. They were popular with staff and easy to deploy. They were also, for the most part, invisible in the income statement. A copilot makes an individual faster; it rarely removes a hand-off, a queue or a control step, and it is those that drive cost in bank operations.

That is now changing. The unit of deployment is shifting from the assistant to the agent — software that plans and executes multi-step work across systems, within limits set by people, and hands exceptions to a human supervisor. The early evidence suggests this is where operational value begins to show up, provided banks redesign the work rather than bolt agents onto it.

The adoption–value gap

A 2025 global survey captures the paradox. 88% of respondents say their organisations regularly use AI in at least one business function, and 62% are at least experimenting with AI agents. But only 23% are scaling an agentic system anywhere in the enterprise, only 39% attribute any EBIT impact to AI, and roughly 6% qualify as high performers whose AI programmes drive 5% or more of EBIT⁴. Nearly two-thirds have not yet begun scaling AI across the enterprise⁴.

The same research points to why. High performers were far more likely to have fundamentally redesigned workflows around AI, rather than layering it onto existing processes⁴. Copilots, by design, leave the workflow intact. Agents force the question.

Where value is showing up

The clearest early results sit in operations that are high-volume, rules-heavy and already well instrumented — the places where banks have long used RPA and offshore teams.

• Reconciliation. On 30 September 2026, Duco reported results from its Pacesetters programme across 12 banks: AI agents working on top of its rules-based reconciliation platform cut overall task time by an average of 76%, with savings of more than 91% in building new reconciliation processes, 84% in optimising them and more than 73% in managing and investigating breaks¹.
• Trade accounting and onboarding. A large US investment bank has spent six months working with embedded Anthropic engineers to co-develop autonomous agents for trade and transaction accounting and for client due diligence and onboarding, according to reporting in February 2026; the bank described itself as in the early stages and expected the agents to significantly reduce the time these core processes take².
• A digital workforce at scale. A US custody bank reports more than 100 ‘digital employees’ deployed through its internal AI platform, each with a distinct persona, credentials and a human supervisor, alongside 99% of staff trained and onboarded to the platform³.

Two features stand out across these examples. First, the agents operate inside a deterministic control framework — a reconciliation engine, an accounting rulebook, a KYC policy — rather than replacing it. Second, a human remains accountable for the outcome. The agent does the assembly work; the person makes the judgement call.

It is also worth being precise about what these results do and do not show. They are early, largely self-reported and concentrated in back- and middle-office processes with clean success criteria. They do not yet demonstrate that agents can safely take customer-facing or credit decisions. But they do show that, where the work is well defined, the productivity step-change is an order of magnitude larger than the incremental gains copilots delivered.

One supervisor, many agents

This is an operating-model change before it is a technology change. In the copilot model, one person uses one assistant to do their existing job a little faster. In the agentic model, one supervisor oversees a team of specialised agents, each with a narrow remit, and spends their time on exceptions, approvals and quality. Published work on financial crime describes human practitioners typically supervising 20 or more AI agents in this way⁵.

Making that work requires the disciplines banks already apply to people and to models, applied to a new kind of worker:

• Identity and entitlements. Each agent has its own credentials and least-privilege access, so its actions are attributable and revocable — the pattern that custody bank describes for its digital employees³.
• Graduated autonomy. Agents start by recommending, then act with approval, and earn wider latitude only on evidence. Regulators' own data show how early the industry is: in the Bank of England and FCA's 2024 survey, 55% of AI use cases had some automated decision-making but only 2% were fully autonomous⁶.
• Evidence by default. Every action, input and rationale is logged, so a supervisor, auditor or examiner can reconstruct why the agent did what it did.
• Measured outcomes. Throughput, accuracy, exception rates and rework are tracked per agent, like any other production process.

What gets in the way

The failure modes are predictable. Agents pointed at fragmented, poorly described data make confident mistakes. Agents deployed without a named owner become nobody's problem until something breaks. And agents bolted onto an unchanged process simply move the queue. The banks reporting results have typically started where the data is structured, the rules are explicit and the outcome is easy to verify — reconciliation breaks, accounting entries, KYC files — and expanded from there.

There is also a cost discipline to learn. Agentic workflows consume far more compute than a single chat interaction, and value depends on unit economics per case handled, not on model benchmarks. Banks that treat agents as production workers — with a cost per task, a quality target and a supervisor — will find the economics easier to manage than those that treat them as an innovation budget line.